SMART: AN INTEGRAL PART OF THE
MULTIPHASE OPTIMIZATION STRATEGY
(MOST)
Presented by Linda M. Collins, Ph.D.
The Methodology Center
Penn State
Presented at the 6th Annual Conference on Statistical Issues in Clinical Trials
Center for Clinical Epidemiology and Biostatistics
University of Pennsylvania
April 17, 2013
Collaborators
• Inbal (Billie) Nahum-Shani
• Daniel Almirall
• Susan A. Murphy
All at University of Michigan
and
• William Pelham, Florida International University
2
Outline
• Definitions
• Development of behavioral interventions: business as
usual
• Development of behavioral interventions: the
Multiphase Optimization Strategy (MOST)
• Using a SMART for optimization
• Ongoing research, open areas and future directions
3
Definition: behavioral intervention
• A program aimed at modifying behavior for the purpose
of treating or preventing disease, promoting health,
and/or enhancing well-being.
• Examples:
—Clinic-based smoking cessation
—Weight loss/management program
—School-based drug abuse prevention
• Note that according to this definition, most behavioral
interventions are treatment packages made up of
multiple components.
4
Definition: Intervention component
• Any aspect of an intervention that can be separated out
for study
—Parts of intervention content
• e.g. : topics in a drug abuse prevention curriculum
—Features that promote compliance/adherence
• e.g.: use of mems caps on medication
—Features aimed at improving fidelity
• e.g.: 800 number for program delivery staff to call with
questions
5
Definition: Intervention component
• Any aspect of an intervention that can be separated out
for study
—A decision rule in a dynamic treatment regime
• e.g.: “If more than one heavy drinking day in a two-week
period, increase clinic visits to three per week”
—A tailoring variable in a dynamic treatment regime
• e.g.: “If more than one heavy drinking day in a two-week
period, increase clinic visits to three per week”
—The review interval in a dynamic treatment regime
• e.g.: “If more than one heavy drinking day in a two-week
period, increase clinic visits to three per week”
6
Definition: Intervention component
• Any aspect of an intervention that can be separated out
for study
—The sequencing of components can itself be considered a
component, particularly in a dynamic treatment regime
• e.g. what should be tried first in treatment of ADHD,
behavioral modification or medication?
7
More about intervention components
• Can impact effectiveness, efficiency, economy, burden,
scalability, etc.
• Some may be pharmaceutical
• Can be defined at any level: individual, family, school,
etc.
8
Outline
• Definitions
• Development of behavioral interventions: business as
usual
• Development of behavioral interventions: the
Multiphase Optimization Strategy (MOST)
• Using a SMART for optimization
• Ongoing research, open areas and future directions
9
How behavioral inventions are typically
developed
• Intervention components are chosen based on scientific
theory, clinical experience, etc.
• Combined into a treatment package
• Package is evaluated via a randomized controlled trial
(RCT)
• The treatment package approach
10
Treatment package approach
component
component
component
component
component
Behavioral
intervention
Evaluation
via RCT
11
What’s wrong with evaluating a
treatment package via an RCT?
Absolutely nothing!
12
The RCT is best suited for
• Determining whether a treatment package performs
better than
—A control or comparison group
—An alternative intervention
13
Treatment package approach
component
component
component
component
component
Behavioral
intervention
Evaluation
via RCT
14
The RCT does not tell us
An RCT that finds a significant effect DOES NOT tell us
• Which components are making positive contributions to
overall effect
• Whether the inclusion of one component has an impact
on the effect of another
• Whether a component’s contribution offsets its cost
• Whether all the components are really needed
• How to make the intervention more effective, efficient,
and scalable
15
What the RCT does not tell us
An RCT that finds a non-significant effect DOES NOT tell
us
• Whether any components are worth retaining
• Whether one component had a negative effect that
offset the positive effect of others
• Specifically what went wrong and how to do it better
the next time
16
Reliance on the treatment package
approach has encouraged…
…stuffing the behavioral intervention with many
components to get a significant effect
…downplaying considerations such as
efficiency
cost-effectiveness
time-effectiveness
participant burden
scalability
17
Reliance on the treatment package
approach has encouraged…
…focusing primarily on attaining statistical significance
…paying insufficient attention to meeting clinically
meaningful criteria
18
What’s the alternative?
• When engineers build products they take an approach
that is
—Systematic
—Efficient
—Focused on the clear objective of optimizing the product
• We can borrow ideas from engineering…
• … and build optimized behavioral interventions
19
Outline
• Definitions
• Development of behavioral interventions: business as
usual
• Development of behavioral interventions: the
Multiphase Optimization Strategy (MOST)
• Using a SMART for optimization
• Ongoing research, open areas and future directions
20
The Multiphase Optimization Strategy
(MOST)
• An engineering-inspired framework for development,
optimization, and evaluation of behavioral interventions
• Using MOST it is possible to engineer a behavioral
intervention to meet a specific optimization criterion
• Three phases: Preparation, Optimization, Evaluation
Collins, Murphy, Nair, & Strecher, 2005; Collins, Murphy, & Strecher, 2007; Collins, Baker, Mermelstein,
Piper, Jorenby, Smith, Schlam, Cook, & Fiore, 2011
21
MOST: Preparation, Optimization,
Evaluation
• Preparation
—Purpose: to lay groundwork for optimization
• Review prior research, take stock of clinical experience,
conduct secondary analyses, etc.
• Derive theoretical model
• Select intervention components to examine
• Conduct pilot/feasibility work
• Identify clearly operationalized optimization criterion
22
Definition: Optimization
• “The process of finding the best possible solution…
subject to given constraints” (The Concise Oxford
Dictionary of Mathematics)
—Optimized does not mean best in an absolute or ideal sense
—Instead, realistic because it includes constraints
• Optimization always involves a clearly stated
optimization criterion
23
Selecting an optimization criterion
• Your definition of “best possible, given constraints”
• This is the goal you want to achieve
• Implicit constraint: the set of components under
consideration
24
One possible optimization criterion:
• Efficient intervention with no “dead wood”
—Select higher/more intense/more expensive level only when
empirically demonstrated to be more effective
• CONSIDER a school-based drug abuse prevention
program.
—Suppose: So that they can justify the use of classroom time
and to reduce waste of resources, the investigators want to be
confident that every component is necessary.
—Achieve this by selecting only active intervention components.
25
This optimization criterion for a
dynamic treatment regime…
• Efficient intervention with no “dead wood”
—Tailoring variables and decision rules that produce the most
effective dynamic treatment regime
—Select higher/more intense/more expensive tailoring variables
and decision rules only when empirically demonstrated to be
more effective
• Will go over an example later
26
Another possible optimization criterion
• Most effective intervention that can be delivered for ≤
some $$
• CONSIDER a dynamic smoking cessation intervention.
—Suppose: Insurers say they will pay for a smoking cessation
intervention that costs no more than $500/person to deliver,
including materials, pharmaceuticals, and staff time.
—Achieve this by identifying a set of tailoring variables and
decision rules that represents the most effective intervention
that can be delivered for ≤ $500.
27
Other possible optimization criteria
• Most effective intervention that can be delivered in ≤
some amount of time
• Cost-effectiveness
• A criterion based on a combination of cost and time
• Most effective without exceeding a specified level of
participant burden
• Or any other relevant criterion
28
MOST: Preparation, Optimization,
Evaluation
• Preparation
—Purpose: to lay groundwork for optimization
• Review prior research, take stock of clinical experience,
conduct secondary analyses, etc.
• Derive theoretical model
• Select intervention components to examine
• Conduct pilot/feasibility work
• Identify optimization criterion
29
MOST: Preparation, Optimization,
Evaluation
• Optimization
—Objective: To form a treatment package that meets the
optimization criterion
• Collect and analyze empirical data on performance of
individual intervention components
— Efficient randomized experimentation
» e.g. fractional factorial, SMART
— Augment by secondary analyses
• Based on information gathered, select components and
levels that meet optimization criterion.
30
MOST: Preparation, Optimization,
Evaluation
• Evaluation
—Objective: To establish whether the optimized intervention
has a statistically significant effect compared to a control or
alternative intervention
• Conduct a randomized clinical trial
31
Treatment package approach
component
component
component
component
component
Behavioral
intervention
Evaluation
via RCT
32
Treatment package approach
component
component
component
component
component
Evaluation
via RCT
33
Multiphase Optimization Strategy
(MOST)
component
component
component
component
component
component
component
Empiricallybased
optimization
Optimized
behavioral
intervention
Evaluation
via RCT
component
34
Multiphase Optimization Strategy
(MOST)
component
component
component
component
component
component
component
Empiricallybased
optimization
Optimized
behavioral
intervention
Evaluation
via RCT
component
35
Multiphase Optimization Strategy
(MOST)
component
component
component
component
component
component
component
Empiricallybased
optimization
Optimized
behavioral
intervention
Evaluation
via RCT
component
36
Evaluation and optimization:
Both important;
not the same thing.
Optimization:
Is the intervention
the best possible,
given constraints?
Evaluation:
Is the intervention’s effect
statistically significant?
No
Yes
No
May wish to
optimize to
improve effect
size
Intervention can
probably be
improved
Yes
Different
intervention
strategy needed
What we should
be aiming for
Outline
• Definitions
• Development of behavioral interventions: business as
usual
• Development of behavioral interventions: the
Multiphase Optimization Strategy (MOST)
• Using a SMART for optimization
• Ongoing research, open areas and future directions
38
Example
• Dynamic treatment regime for children with ADHA
• W. Pelham is the intervention scientist
• Optimization criterion: Efficient dynamic treatment
regime with no “dead wood”
39
Dynamic treatment regimes for children
with ADHD
• Two approaches to treatment of ADHD
—Behavior modification (BMOD)
—Medication
• They can be combined into a dynamic treatment regime
• What is the best set of decision rules?
40
Dynamic treatment regimes for children
with ADHD
• Research questions that must be addressed for
optimization:
—Is it better to start with BMOD or medication?
• Note: BMOD much more expensive
—For those who do not respond to initial treatment, is it better
to
• Remain with the same strategy but enhance
• Augment with the other strategy
—What is the best overall strategy?
41
SMART
Continue Med (SG1)
Response
Medication
Enhance (SG2)
Non-Response
R
Augment (SG3)
R
Response
Continue BMOD (SG4)
BMOD
Enhance (SG5)
Non-Response
R
Augment (SG6)
42
Questions we can address with SMART
• Is it better to start with BMOD or MED?
• (SG1+SG2+SG3) vs. (SG4+SG5+SG6)
• Medication vs. BMOD
—Averaging over subsequent treatment
Continue (SG1)
Response
Enhance (SG2)
MED
Non-Response
R
Augment (SG3)
R
Response
Continue (SG4)
BMOD
Enhance (SG5)
Non-Response
R
Augment (SG6)
43
Questions we can address with SMART
• Is it better to Enhance or Augment for nonresponders?
• (SG2+SG5) vs. (SG3+SG6)
• Enhance vs. Augment
Continue (SG1)
Response
Enhance (SG2)
MED
Non-Response
R
Augment (SG3)
R
Response
Continue (SG4)
BMOD
Enhance (SG5)
Non-Response
R
Augment (SG6)
44
Questions we can address with SMART
• What is the best overall strategy? There are FOUR embedded here.
Stage 1 = {BMOD},
IF response = {NO}
THEN stage 2 = {AUGMENT}
ELSE continue stage 1
Stage 1 = {MED},
IF response = {NO}
THEN stage 2 = {AUGMENT}
ELSE continue stage 1
Stage 1 = {MED},
IF response = {NO}
THEN stage 2 = {ENHANCE}
ELSE continue stage 1
Enhance (SG2)
MED
Non-Response
Stage 1 = {BMOD},
IF response = {NO}
THEN stage 2 = {ENHANCE}
ELSE continue stage 1
Continue (SG1)
Response
R
Augment (SG3)
R
Response
Continue (SG4)
BMOD
Enhance (SG5)
Non-Response
R
Augment (SG6)
45
After the SMART
• Results of analyses will be used to decide on best
decision rules, i.e. optimized intervention
• This does not tell us whether the optimized intervention
has a statistically significant effect as compared to a
control
• Must move to the Evaluation phase for that
46
Outline
• Definitions
• Development of behavioral interventions: business as
usual
• Development of behavioral interventions: the
Multiphase Optimization Strategy (MOST)
• Using a SMART for optimization
• Ongoing research, open areas and future directions
47
Ongoing research, open areas, and future
directions
• Q-learning (Watkins, 1989; Murphy, 2005)
—Popular method from computer science.
—Regression-based:
• Sequence of regressions
• One regression for each stage
—Rationale:
• Base your decision on what you know up to that point
• Assume best future decisions
Nahum-Shani et al., (2012). Q-learning: A secondary data analysis method for
developing adaptive interventions. Psychological Methods, 17(4), 478-494
48
Ongoing research, open areas, and future
directions
• Building a controller using techniques from engineering
—In much the same way any other controller would be built
—Requires intensive longitudinal data
—Must be able identify system
• Usually will require experimentation
—Potential for m-health interventions
Zafra-Cabeza, A., Rivera, D.E., Collins, L.M., Ridao, M.A., & Camacho, E.F. (2011). A riskbased model predictive control approach to adaptive interventions in behavioral health.
IEEE Transactions on Control Systems Technology, 19, 891-901.
Trail, J.B., Collins, L.M., Rivera, D.E.,Li, R., Piper, M., & Baker, T. (almost in press). Functional
Data Analysis for the system identification of behavioral processes. Psychological
Methods.
Nahum-Shani, I., ISR, University of Michigan
49
Ongoing research, open areas, and future
directions
• Q-learning as a secondary data analysis approach for
data from SMART’s (Nahum-Shani et al., 2012)
• Clarification about optimization criteria, particularly
with dynamic treatment regimes
• Experimental designs that mix adaptive (dynamic) and
fixed components
• More applications of MOST, particularly including
SMARTs
50
For more information:
http://methodology.psu.edu
LMCOLLINS@PSU.EDU
Some Publications
Chakraborty, B., Collins, L.M., Strecher, V., and Murphy, S.A. (2009). Developing multicomponent interventions using
fractional factorial designs. Statistics in Medicine, 28, 2687-2708.
Collins, L.M., Baker, T.B., Mermelstein, R.J., Piper, M.E., Jorenby, D.E., Smith, S.S., Schlam, T.R., Cook, J.W., & Fiore,
M.C. (2011). The Multiphase Optimization Strategy for engineering effective tobacco use interventions.
Annals of Behavioral Medicine, 41, 208-226.
Collins, L.M., Chakraborty, B., Murphy, S.A., & Strecher, V. (2009). Comparison of a phased experimental approach
and a single randomized clinical trial for developing multicomponent behavioral interventions. Clinical Trials,
6, 5-15.
Collins, L.M., Dziak, J.R., & Li, R. (2009). Design of experiments with multiple independent variables: A resource
management perspective on complete and reduced factorial designs. Psychological Methods, 14, 202-224.
Nahum-Shani et al., (2012). Experimental design and primary data analysis for developing adaptive interventions.
Psychological Methods, 17 (4) 457-477.
Nahum-Shani et al., (2012). Q-Learning: A secondary data analysis method for developing adaptive interventions.
Psychological Methods, 17(4), 478-494.
Trail, J.B., Collins, L.M., Rivera, D.E.,Li, R., Piper, M., & Baker, T. (almost in press). Functional Data Analysis for the
system identification of behavioral processes. Psychological Methods.
Zafra-Cabeza, A., Rivera, D.E., Collins, L.M., Ridao, M.A., & Camacho, E.F. (2011). A risk-based model predictive
control approach to adaptive interventions in behavioral health. IEEE Transactions on Control Systems
Technology, 19, 891-901.